test: am of empty patch should not succeed

test: am of empty patch should not succeed

The "git am empty" test uses the construct

	git am empty-file && false || :

which unconditionally returns true.  Use test_must_fail instead, which
also has the benefit of noticing if "git am" has segfaulted.

While at it, tighten the test to check that the diagnostic appears on
stderr and not stdout.

Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>